We’re Ben and Jacob, cofounders of Freestyle (https://freestyle.sh). We’re building a cloud for Coding Agents.
For the first generation of agents it looked like workflows with minimal tools. 2 years ago we published a package to let AI work in SQL, at that time GPT-4 could write simple scripts. Soon after the first AI App Builders started using AI to make whole websites; we supported that with a serverless deploy system.
But the current generation is going much further, instead of minimal tools and basic serverless apps AI can utilize the full power of a computer (“sandbox”). We’re building sandboxes that are interchangeable with EC2s from your agents perspective, with bonus features:
1. We’ve figured out how to fork a sandbox horizontally without more than a 400ms pause in it. That's not forking the filesystem, we mean forking the whole memory of it. If you’re half way down a browser page with animations running, they’ll be in the same place in all the forks. If you’re running a minecraft server every block and player will be in the same place on the forks. If you’re running a local environment and an error comes up in process that error will be there in all the forks. This works for snapshotting as well, you can save your place and come back weeks later.
2. Our sandboxes start in ~500ms.
Demo: https://www.loom.com/share/8b3d294d515442f296aecde1f42f5524
Compared with other sandboxes, our goal is to be the most powerful. We support full Linux + hardware-virtualization, eBPF, Fuse, etc. We run full Debian with multiple users and we use a systemd init instead of runc. Whatever your AI expects to work on debian should work on these vms, and if it doesn’t send a bug report.
In order to make this possible, we’ve moved to our own bare metal racks. Early in our testing we realized that moving VMs across cloud nodes would not have acceptable performance properties. We asked Google Cloud and AWS for a quote on their bare metal nodes and found that the monthly cost was equivalent to the total cost of the hardware so we did that.
Our goal is to build the necessary infrastructure to replicate the human devloop on the massively multi-tenant scale of AI, so these VMs should be as powerful as the ones you’re used to, while also being available to provision in seconds.
Would love to understand how you compare to other providers like Modal, Daytona, Blaxel, E2B and Vercel. I think most other agent builders will have the same question. Can you provide a feature/performance comparison matrix to make this easier?
I'm working on an article deep diving into the differences between all of us. I think the goal of Freestyle is to be the most powerful and most EC2 like of the bunch.
Daytona runs on Sysbox (https://github.com/nestybox/sysbox) which is VM-like but when you run low level things it has issues.
Modal is the only provider with GPU support.
I haven't played around with Blaxel personally yet.
E2B/Vercel are both great hardware virtualized "sandboxes"
Freestyle VMS are built based on the feedback our users gave us that things they expected to be able to do on existing sandboxes didn't work. A good example here is Freestyle is the only provider of the above (haven't tested blaxel) that gives users access to the boot disk, or the ability to reboot a VM.
And fly.io sprites
Fly.io sprites is the most similar to us of the bunch. They do hardware virtualization as well, have comparable start times and are full Linux. What we call snapshots they call checkpoints.
The big pros of Sprites over us is their advanced networking stack and the Fly.io ecosystem. The big cons are that Sprites are incredibly bare bones — they don't have any templating utilities. I've also heard that Sprites sometimes become unavailable for extended periods of time.
The big pros of Freestyle over Sprites is fork, advanced templating, and IMO a better debugging experience because of our structure.
Thanks for the thoughtful response. I'm predominantly a self-hoster, but I think your product makes a lot of sense for a wide variety of users and businesses. I'm excited to try out freestyle!
Self hosting can be doable for constant small/medium size workloads
You can handroll a lot with: https://github.com/nestybox/sysbox?tab=readme-ov-file https://gvisor.dev https://github.com/containers/bubblewrap?tab=readme-ov-file
For hardware virtualized machines it much harder but you can do it via: https://github.com/firecracker-microvm/firecracker/ https://github.com/cloud-hypervisor/cloud-hypervisor
Freestyle/other providers will likely provide better debugging experience but thats something you can probably get past for a lot of workloads.
The time when you/anyone should think about Freestyle/anyone is when the load spikes/the need to create hundreds of VMs in short spikes shows up, or when you're looking for some of the more complex feature sets any given provider has built out (forks, GPUs, network boundaries, etc).
I also highly recommend self hosting anything you do outside of your normal VPC. Sandboxes are the biggest possible attack surface and it is a feature of us that we're not in your cloud; If we mess up security your app is still fine.
This is what I do (my project) for self hosting on a VPS/server:
https://GitHub.com/jgbrwn/vibebin
Also I'm a huge proponent of exe.dev
Obviously your service/approach is different than exe, more like sprites but like you said more targeted/opinionated to AI coding/sandboxing tasks it looks like. Interesting space for sure!
I've been building an open-source, self-hostable Firecracker orchestrator for the past month: https://github.com/sahil-shubham/bhatti (https://bhatti.sh)
Still WIP, but the core works — three rootfs tiers (minimal Ubuntu, headless Chromium with CDP, Docker-in-VM), OCI image support (pull any Docker image), automatic thermal management (idle VMs pause then snapshot to disk, wake transparently on next API call), per-user bridge networking with L2 isolation, named checkpoints, persistent volumes, and preview URLs with auto-wake.
Fair warning: the website is too technical and the docs are mostly AI-generated, both being actively reworked. But I've been running it daily on a Hetzner server for my AI agents' browser automation, and deploy previews.
I'd love any feedback if you want to go ahead and try it yourself
sprites have weird lately, i think fly.io is having trouble with capacity in various locations.
is the experience similar? can i just get console to one machine, work for a bit, logout. come back later, continue?
how does i cost work if i log into a machine and do nothing on it? just hold the connection.
This will just work on us.
We do auto suspend depending on your configured timeout. We'll pause your VM and when you come back the processes will be in the exact same state as when you left.
I'd also be interested in a comparison with exe.dev which I'm currently using.
Exe.dev is a individual developer oriented service. Freestyle is more oriented at platforms building the next exe.dev.
Thats why our pricing is usage based and we have a much larger API surface.
Is it possible to run a Kubernetes cluster inside one? (E.g. via KIND.)
If so, we'd very much like to test this. We make extensive use of Claude Code web but it can't effectively test our product inside the sandbox without running a K8s cluster
Your UI design is really nice.
Non open source and non local SAAS sandboxes are offensive to even try to launch. No one needs this and the only customers will be vibe coders who just don't know any better. There are teams building actual sandboxes like smolmachines, podman, colima and mre. At least be honest and put the virtualisation tech you are using as well as that its closed source SAAS on the landing page to safe people time.
Our users are platforms, and many of the best already build on us.
Self hosting is a valuable feature but our technology is unfriendly to small nodes — it will not work on consumer hardware. Many of the optimizations we spend our time on only seriously kick in above 2TB of storage and above 500GB of RAM.
Is this similar to https://instavm.io/?
Never tried them, I think the weird thing about VM providers is the difference really all is in the execution. These guys seem great in concept but I don’t know enough about how they properly work.
I’m super interested since it seems like you have given everything a lot of thought and effort but I am not sure I understand it.
When I’m thinking of sandboxes, I’m thinking of isolated execution environments.
What does forking sandboxes bring me? What do your sandboxes in general bring me?
Please take this in the best possible way: I’m missing a use case example that’s not abstract and/or small. What’s the end goal here(
So isolation is correct. Forking a sandbox gives you multiple exact duplicates of isolated environments.
When your coding agent has 10 ideas for what to do, to evaluate them correctly it needs to be able to evaluate them in isolation.
If you're building a website testing agent and halfway down a website, with a form half filled out a session ongoing, etc and it realizes it wants to test 2 things in isolation, forking is the only way.
We also envision this powering the next generation of devcycles "AI Agent, go try these 10 things and tell me which works best". AI forks the environment 10 times, gets 10 exact copies, does the thing in each of them, evaluates it, then takes the best option.
Yep I can see this especially when the agent is spinning up test servers/smokes and you don't want those conflicting. How do we reconcile all the potential different git hashes though, upstream I guess etc (this might be an easy answer and I'm not super proficient with git so forgive)
So we recommend branch per fork, merge what you like.
You have to change the branch on each fork individually currently and thats unlikely to change in the short term due to the complexity of git internals, but its not that hard to do yourself `git checkout -b fork-{whateverDiscriminator}`
Agreed, the thing I'd be most interested in is the isolated execution environment you mentioned. Agents running autopilot are powerful. Agents running unsupervised on a machine with developer permissions and certificates where anything could influence the agent to act on an attacker's behalf is terrifying
I recommend running the agent harness outside of the computer. The mental model I like to use is the computer is a tool the agent is using, and anything in the computer is untrusted.
I would recommend not giving an agent the full run of any computing environment. Do handle fine grained internet access controls and credential injection like OpenShell does?
I used to believe this, but I think the next generation of agents is much more autonomous and just needs a computer.
The work of a developer is open ended, so we use a computer for it. We don't try to box developers into small granular screwdrivers for each small thing.
Thats whats coming to all agents, they might want to run some analysis with python, want to generate a website/document in typescript, and might want to store data in markdown files or in MongoDB. I expect them to get much more autonomous and with that to end up just needing computers like us.
The problem is the agent, which should be treated untrusted. The computer isn’t the problem
Kind of. The chat logs of the agent are trustworthly, as should any telemetry you have on it or coming out of the VM. Its behavior should be treated as probabilistic and therefore untrustworthly.
Wow, forking memory along with disk space this quickly is fascinating! That's something that I haven't seen from your competitors.
If the machine can fork itself, it could allow for some really neat auto-forking workflows where you fuzz the UI testing of a website by forking at every decision point. I forget the name of the recent model that used only video as its latent space to control computers and cars, but they had an impressive demo where they fuzzed a bank interface by doing this, and it ended up with an impressive number of permutations of reachable UI states.
That’s what I’m hoping for!
The technical challenges in getting memory forking to deliver those sub-second start and fork times are significant. I've seen the pain of trying to achieve that level of state transfer and rapid provisioning. While "EC2-like" gets the point across for many, going bare metal reveals the practical limits of cloud virtualization for high-performance, complex workloads like these. It shows a real understanding of where cloud abstraction helps and where it just adds overhead.
The cost argument for owning the hardware for this specific use case also makes sense, considering the scale these agent environments will demand. Also worth noting, sandboxes are effectively an open attack surface; architecting them not to be in your main VPC is a sound security decision from the start.
Congratulations on the launch! Will definitely test this out.
This is awesome - the snapshotting especially is critical for long running agents. Since we run agents in a durable execution harness (similar to Temporal / DBOS) we needed a sandboxing approach that would snapshot the state after every execution in order to be able to restore and replay on any failure.
We ended up creating localsandbox [0] with that in mind by using AgentFS for filesystem snapshotting, but our solution is meant for a different use case than Freestyle - simpler FS + code execution for agents all done locally. Since we're not running a full OS it's much less capable but also simpler for lots of use cases where we want the agent execution to happen locally.
The ability to fork is really interesting - the main use case I could imagine is for conversations that the user forks or parallel sub-agents. Have you seen other use cases?
Deterministic testing of edge cases. It can be really hard to recreate weird edge cases of running services, but if you can create them we can snapshot them exactly as they are.
I built something like this at work using plain Docker images. Can you help me understand your value prop a little better?
The memory forking seems like a cool technical achievement, but I don't understand how it benefits me as a user. If I'm delegating the whole thing to the AI anyway, I care more about deterministic builds so that the AI can tackle the problem.
So first MicroVM != Container, and container is not a secure isolation system. I would not run untrusted containers on your nodes without extra hardening.
The memory forking was originally invented because for AI App Builders and first response driven applications its extremely important that they are instant (difference between running bun dev and the dev server already being running).
However its much more generally applicable, Postgres is a great example of this. You can't fork the filesystem under postgres and get consistency. Same thing with a browser state, a weird server state, or anything that exists in memory. The memory forking gives a huge performance boost while snapshotting whats actually going on at one instant.
I think one of the very few who actually support ebpf & xdp, which you do need when you're building low level stuff. + the bare metal setup is like out of the world lol.
Tx it took a lot of work lol
Any ideas for locking down remote access from an untrusted VM? Cloudflare has object-based capabilities and some similar thing might be useful to let a VM make remote requests without giving it API keys. (Keys could be exfiltrated via prompt injection.)
So we have there are 3 solutions to this, Freestyle supports 2 of them: 1. Freestyle supports multiple linux users. All linux users on the VM are locked down, so its safe to have a part of the vm that has your secret keys/code that the other parts cannot access. 2. A custom proxy that routes the traffic with the keys outside 3. We're working on a secrets api to intercept traffic and inject keys based on specific domains and specific protocols starting with HTTP Headers, HTTP Git Authentication and Postgres. That'll land in a few weeks.
Cool! I've been using your API for running sandboxed JS. Nice to see you also support VMs now.
> we mean forking the whole memory of it
We're using copy on write with the memory itself. Fork time is completely decoupled from the size of the machine.
Creating snapshots takes a 2-4 second interruption in the VM due to sheer IO that we didn't want here.
Whats especially cool about this approach is not only is fork time O(1) with respect to machine size, but its also O(1) with respect to the amount of forks.
It doesn't seem very easy to calculate how much it would cost per month to keep a mostly-idle VM running (for example, with a personal web app). The $20/month plan from exe.dev seems more hobbyist-friendly for that. Maybe that's not the intended use, though?
We're not going after hobbyists. We're building the platform for companies like exe.dev to build on. Thats why its all usage based.
That said, our $50 a month plan can be used as an individual for your coding agents, but I wouldn't recommend it.
Ooof, if you are the middleman platform then it's sure gonna get expensive for the end user
> The $20/month plan from exe.dev seems more hobbyist-friendly for that. Maybe that's not the intended use, though?
And you can go even below that by self-hosting it yourself with a very cheap Hetzner box for $2 or $5.
Can you start up multiple VM's easily on a Hetzner box?
how many seconds to provision are we talking about here? 1 sec vs 60 is a dealbreaker for me, some clarity on that would be nice.
500ms. Less than 1 second. We're aiming to get that down to 200ms in the next 3 months.
Honestly never considered the forking use case; but it makes a ton of sense when explained
Congrats on the launch. This is cool tech
Congrats Ben and Jacob!
Interesting!
We're working on a similar solution at UnixShells.com [1]. We built a VMM that forks, and boots, in < 20ms and is live, serving customers! We have a lot of great tools available, via MIT, on our github repo [2] as well!
Can your service scale ram? like the way docker desktop does. Manual is fine.
yep you can choose ram + disk + cpu size
It's hard to tell what this is or how it compares to other things that are out there, but what I latched onto is this:
> Freestyle is the only sandbox provider with built-in multi-tenant git hosting — create thousands of repos via API and pair them directly with sandboxes for seamless code management. On top of that, Freestyle VMs are full Linux virtual machines with nested virtualization, systemd, and a complete networking stack, not containers.
It makes me think of the git automation around rigs in Gas Town: https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16d...
Edit: I realize the Loom is a way to look at it. Loom interrupted me twice and I almost skipped it. However it gave me a better idea of what it does, it "invents" snapshotting and restoring of VMs in a way that appears faster. That actually makes sense and I know it isn't that hard to do with how VMs work and that it greatly benefits from having only part of the VM writable and having little memory used (maybe it has read-only memory too?).
So the snapshotting tech is actually 100% independent of Git.
Git is useful for branching vs forking (IE you can't merge two VM forks back together), but all the tech I showed in the Loom exists independently from Git.
The hard part of it was making the VM large and powerful while making snapshotting/forking instant, which required a lot of custom VMM work.
Your pricing page is broken
Reviewing this now. our public pricing at www.freestyle.sh/pricing seems to be working, can you point me in a more specific direction?
dumb question. none of these protect your from prompt injection. yes?
no, but the goal of these is if you are faced with prompt injection the worst case scenario is the AI uses that computer badly.
unless i am misundestanding. not sure how this computer prevents secrets from my gmail leaking. thats the worst case.
If you put your gmail credentials into a VM that an AI Agent dealing with untrusted prompts has access to they should be treated as leaked and be disabled immediately.
However, if you don't put your administrative credentials inside of the VM and treat it as an unsafe environment you can safely give it minimal permissions to access specific things that it needs and using that access it can perform complex tasks.
i am talking about this . not my gmail credentials.
https://simonwillison.net/2024/Mar/5/prompt-injection-jailbr...
I have so many interesting problems on Ai, sandboxing isn't one of them. It's a pointless excercise yet disproportionately so many people love to to do this. Probably because sandboxing doesn't feel as magic as Agents itself and more like the old times of "traditional" software development.
It is a mostly pointless exercise if the goal is trying to contain negative impact of AI agents (e.g. OpenClaw).
It is a very necessary building block for many common features that can be steered in a more deterministic way, e.g. "code interpreter" feature for data analysis or file creation like commonly seen in chat web UIs.
Believe it or not, once you start working for a regulated industry, it is all you would ever think of. There, people don't care if you are vibing with the latest libraries and harnesses or if it's magic, they care that the entire deployment is in some equivalent of a Faraday cage. Plus, many people just don't appreciate it when their agents go rm -rf / on them.
Yeah, idk I guess it’s interesting if you are an engineer looking for something to do,
But like I see multiple sandbox for agents products a week. Way too saturated of a market
I disagree (as a sandboxing company).
With respect to the market, every single sandbox sucks. I'm not gonna shit talk competitors but there is not a good sandboxing platform out there yet — including me — compared to where we'll be in 6 months.
We've heard all the platforms have consistent uptime, feature completeness, networking and debugging issues. And in our own platform we're not 1/10ths of the way through solving the requests we've gotten.
Next generation of Agents needs computers, and those computers are gonna look really different than "sandboxes" do today.
I don't think you're wrong, but if you really want to really re-think the approach, building an orchestration layer for Firecracker like every other company in the space is doing is probably not it.